Abstract: Image similarity learning is to predict whether two images match by learning image content information through the network. Until now, the improved variant network based on convolutional neural network has effectively improved the learning efficiency. However, because the extracted features are relatively single and cannot accurately describe the image features, the similarity learning efficiency is low. To this end, a two-branch network based on a convolutional neural network structure is proposed. The network is a left and right branch network, the structure is the same but the weights are not shared, and the network input is a two-branch input. Firstly, the single channel image features are extracted by the left and right branch networks respectively. Then the features fused through the feature fusion layer. Finally, the fusion feature is directly input into the fully connected layer for similarity learning. It not only improves the feature diversity of the extracted images, but also speeds up the training of the model. Comparative experiments were carried out on the chip card slot image dataset taken by the laboratory industrial camera. The experimental results show that compared with other models, the proposed model has strong network learning ability and model generalization ability, and the accuracy rate is 97.96%.

Key words: image similarity learning, convolutional neural network, two-branch network, weight not shared, feature fusion